ADR-008 — yolo/v1 task profile: single-shot only, no callback bridge, model-family-named

Status: proposed (2026-08-30, spec v0.2). Resolves the "detection" portion of OQ12;
embedding and tts remain open.
Context: Three independent, real implementations exist for YOLO-family object detection across this ecosystem — an Android/ONNX Runtime app, a proven on-device NPU pipeline (vendor NPU/IREE), and a SKaiNET-DSL-native graph (currently a zero-weight skeleton, unverified against IREE) — none sharing a common public interface. Unlike ASR, none of the real cartridge candidates behind this profile stream: all three capture one full image and infer once. There is also no closed, app-internal precedent to extract from (no app-internal existing contract for detection), unlike asr/v1’s extraction from the first consumer app’s voice-engine interface (ADR-005).
Decision: Add `yolo/v1 to cartridge-task-profiles as a single, Flow-free, session-free contract: one YoloEngine.infer(image: YoloImage): YoloResult suspend call, mirroring the C ABI’s ctg_infer exactly (one whole input, one terminal result). No callback bridge ships in v1 — unlike asr/v1’s two-layer design, there is no known callback-structured consumer for detection today, so the bridge would be speculative machinery; add one if and when a real consumer needs it. The profile is deliberately named after the model family (`yolo) rather than the task (detection), departing from asr/v1’s task-based naming — the real cartridges behind it are consistently YOLO-family, and a differently-shaped detector, if one ever appears, can get its own profile rather than forcing `yolo/v1 to abstract over an architecture it was never validated against.
Consequences: a streaming/video variant, if ever needed, is a new profile version (yolo/v2) rather than a capability flag on this one — yolo/v1 does not attempt to hide a streaming capability the way asr/v1 must. The three candidate implementations named above can each adapt to this one interface without redesigning it around any single one’s current shape. If a future non-YOLO detector needs the same shape, revisit the model-family-named precedent this ADR sets rather than silently reusing yolo/v1 for it.